lustre/ptlrpc/niobuf.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, 2013, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_RPC
  38 #include <obd_support.h>
  39 #include <lustre_net.h>
  40 #include <lustre_lib.h>
  41 #include <obd.h>
  42 #include <obd_class.h>
  43 #include "ptlrpc_internal.h"
  44
  45 /**
  46  * Helper function. Sends \a len bytes from \a base at offset \a offset
  47  * over \a conn connection to portal \a portal.
  48  * Returns 0 on success or error code.
  49  */
  50 static int ptl_send_buf (lnet_handle_md_t *mdh, void *base, int len,
  51                          lnet_ack_req_t ack, struct ptlrpc_cb_id *cbid,
  52                          struct ptlrpc_connection *conn, int portal, __u64 xid,
  53                          unsigned int offset)
  54 {
  55         int              rc;
  56         lnet_md_t         md;
  57         ENTRY;
  58
  59         LASSERT (portal != 0);
  60         LASSERT (conn != NULL);
  61         CDEBUG (D_INFO, "conn=%p id %s\n", conn, libcfs_id2str(conn->c_peer));
  62         md.start     = base;
  63         md.length    = len;
  64         md.threshold = (ack == LNET_ACK_REQ) ? 2 : 1;
  65         md.options   = PTLRPC_MD_OPTIONS;
  66         md.user_ptr  = cbid;
  67         md.eq_handle = ptlrpc_eq_h;
  68
  69         if (unlikely(ack == LNET_ACK_REQ &&
  70                      OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_ACK, OBD_FAIL_ONCE))){
  71                 /* don't ask for the ack to simulate failing client */
  72                 ack = LNET_NOACK_REQ;
  73         }
  74
  75         rc = LNetMDBind (md, LNET_UNLINK, mdh);
  76         if (unlikely(rc != 0)) {
  77                 CERROR ("LNetMDBind failed: %d\n", rc);
  78                 LASSERT (rc == -ENOMEM);
  79                 RETURN (-ENOMEM);
  80         }
  81
  82         CDEBUG(D_NET, "Sending %d bytes to portal %d, xid "LPD64", offset %u\n",
  83                len, portal, xid, offset);
  84
  85         rc = LNetPut (conn->c_self, *mdh, ack,
  86                       conn->c_peer, portal, xid, offset, 0);
  87         if (unlikely(rc != 0)) {
  88                 int rc2;
  89                 /* We're going to get an UNLINK event when I unlink below,
  90                  * which will complete just like any other failed send, so
  91                  * I fall through and return success here! */
  92                 CERROR("LNetPut(%s, %d, "LPD64") failed: %d\n",
  93                        libcfs_id2str(conn->c_peer), portal, xid, rc);
  94                 rc2 = LNetMDUnlink(*mdh);
  95                 LASSERTF(rc2 == 0, "rc2 = %d\n", rc2);
  96         }
  97
  98         RETURN (0);
  99 }
 100
 101 static void mdunlink_iterate_helper(lnet_handle_md_t *bd_mds, int count)
 102 {
 103         int i;
 104
 105         for (i = 0; i < count; i++)
 106                 LNetMDUnlink(bd_mds[i]);
 107 }
 108
 109 #ifdef HAVE_SERVER_SUPPORT
 110 /**
 111  * Prepare bulk descriptor for specified incoming request \a req that
 112  * can fit \a npages * pages. \a type is bulk type. \a portal is where
 113  * the bulk to be sent. Used on server-side after request was already
 114  * received.
 115  * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
 116  * error.
 117  */
 118 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
 119                                               unsigned npages, unsigned max_brw,
 120                                               unsigned type, unsigned portal)
 121 {
 122         struct obd_export *exp = req->rq_export;
 123         struct ptlrpc_bulk_desc *desc;
 124
 125         ENTRY;
 126         LASSERT(type == BULK_PUT_SOURCE || type == BULK_GET_SINK);
 127
 128         desc = ptlrpc_new_bulk(npages, max_brw, type, portal);
 129         if (desc == NULL)
 130                 RETURN(NULL);
 131
 132         desc->bd_export = class_export_get(exp);
 133         desc->bd_req = req;
 134
 135         desc->bd_cbid.cbid_fn  = server_bulk_callback;
 136         desc->bd_cbid.cbid_arg = desc;
 137
 138         /* NB we don't assign rq_bulk here; server-side requests are
 139          * re-used, and the handler frees the bulk desc explicitly. */
 140
 141         return desc;
 142 }
 143 EXPORT_SYMBOL(ptlrpc_prep_bulk_exp);
 144
 145 /**
 146  * Starts bulk transfer for descriptor \a desc on the server.
 147  * Returns 0 on success or error code.
 148  */
 149 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc)
 150 {
 151         struct obd_export        *exp = desc->bd_export;
 152         struct ptlrpc_connection *conn = exp->exp_connection;
 153         int                       rc = 0;
 154         __u64                     xid;
 155         int                       posted_md;
 156         int                       total_md;
 157         lnet_md_t                 md;
 158         ENTRY;
 159
 160         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_PUT_NET))
 161                 RETURN(0);
 162
 163         /* NB no locking required until desc is on the network */
 164         LASSERT(desc->bd_md_count == 0);
 165         LASSERT(desc->bd_type == BULK_PUT_SOURCE ||
 166                 desc->bd_type == BULK_GET_SINK);
 167
 168         LASSERT(desc->bd_cbid.cbid_fn == server_bulk_callback);
 169         LASSERT(desc->bd_cbid.cbid_arg == desc);
 170
 171         /* NB total length may be 0 for a read past EOF, so we send 0
 172          * length bulks, since the client expects bulk events.
 173          *
 174          * The client may not need all of the bulk XIDs for the RPC.  The RPC
 175          * used the XID of the highest bulk XID needed, and the server masks
 176          * off high bits to get bulk count for this RPC. LU-1431 */
 177         xid = desc->bd_req->rq_xid & ~((__u64)desc->bd_md_max_brw - 1);
 178         total_md = desc->bd_req->rq_xid - xid + 1;
 179
 180         desc->bd_md_count = total_md;
 181         desc->bd_failure = 0;
 182
 183         md.user_ptr = &desc->bd_cbid;
 184         md.eq_handle = ptlrpc_eq_h;
 185         md.threshold = 2; /* SENT and ACK/REPLY */
 186
 187         for (posted_md = 0; posted_md < total_md; xid++) {
 188                 md.options = PTLRPC_MD_OPTIONS;
 189
 190                 /* NB it's assumed that source and sink buffer frags are
 191                  * page-aligned. Otherwise we'd have to send client bulk
 192                  * sizes over and split server buffer accordingly */
 193                 ptlrpc_fill_bulk_md(&md, desc, posted_md);
 194                 rc = LNetMDBind(md, LNET_UNLINK, &desc->bd_mds[posted_md]);
 195                 if (rc != 0) {
 196                         CERROR("%s: LNetMDBind failed for MD %u: rc = %d\n",
 197                                exp->exp_obd->obd_name, posted_md, rc);
 198                         LASSERT(rc == -ENOMEM);
 199                         if (posted_md == 0) {
 200                                 desc->bd_md_count = 0;
 201                                 RETURN(-ENOMEM);
 202                         }
 203                         break;
 204                 }
 205                 /* Network is about to get at the memory */
 206                 if (desc->bd_type == BULK_PUT_SOURCE)
 207                         rc = LNetPut(conn->c_self, desc->bd_mds[posted_md],
 208                                      LNET_ACK_REQ, conn->c_peer,
 209                                      desc->bd_portal, xid, 0, 0);
 210                 else
 211                         rc = LNetGet(conn->c_self, desc->bd_mds[posted_md],
 212                                      conn->c_peer, desc->bd_portal, xid, 0);
 213
 214                 posted_md++;
 215                 if (rc != 0) {
 216                         CERROR("%s: failed bulk transfer with %s:%u x"LPU64": "
 217                                "rc = %d\n", exp->exp_obd->obd_name,
 218                                libcfs_id2str(conn->c_peer), desc->bd_portal,
 219                                xid, rc);
 220                         break;
 221                 }
 222         }
 223
 224         if (rc != 0) {
 225                 /* Can't send, so we unlink the MD bound above.  The UNLINK
 226                  * event this creates will signal completion with failure,
 227                  * so we return SUCCESS here! */
 228                 spin_lock(&desc->bd_lock);
 229                 desc->bd_md_count -= total_md - posted_md;
 230                 spin_unlock(&desc->bd_lock);
 231                 LASSERT(desc->bd_md_count >= 0);
 232
 233                 mdunlink_iterate_helper(desc->bd_mds, posted_md);
 234                 RETURN(0);
 235         }
 236
 237         CDEBUG(D_NET, "Transferring %u pages %u bytes via portal %d "
 238                "id %s xid "LPX64"-"LPX64"\n", desc->bd_iov_count,
 239                desc->bd_nob, desc->bd_portal, libcfs_id2str(conn->c_peer),
 240                xid - posted_md, xid - 1);
 241
 242         RETURN(0);
 243 }
 244 EXPORT_SYMBOL(ptlrpc_start_bulk_transfer);
 245
 246 /**
 247  * Server side bulk abort. Idempotent. Not thread-safe (i.e. only
 248  * serialises with completion callback)
 249  */
 250 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc)
 251 {
 252         struct l_wait_info       lwi;
 253         int                      rc;
 254
 255         LASSERT(!in_interrupt());           /* might sleep */
 256
 257         if (!ptlrpc_server_bulk_active(desc))   /* completed or */
 258                 return;                         /* never started */
 259
 260         /* We used to poison the pages with 0xab here because we did not want to
 261          * send any meaningful data over the wire for evicted clients (bug 9297)
 262          * However, this is no longer safe now that we use the page cache on the
 263          * OSS (bug 20560) */
 264
 265         /* The unlink ensures the callback happens ASAP and is the last
 266          * one.  If it fails, it must be because completion just happened,
 267          * but we must still l_wait_event() in this case, to give liblustre
 268          * a chance to run server_bulk_callback()*/
 269         mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_count);
 270
 271         for (;;) {
 272                 /* Network access will complete in finite time but the HUGE
 273                  * timeout lets us CWARN for visibility of sluggish NALs */
 274                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 275                                            cfs_time_seconds(1), NULL, NULL);
 276                 rc = l_wait_event(desc->bd_waitq,
 277                                   !ptlrpc_server_bulk_active(desc), &lwi);
 278                 if (rc == 0)
 279                         return;
 280
 281                 LASSERT(rc == -ETIMEDOUT);
 282                 CWARN("Unexpectedly long timeout: desc %p\n", desc);
 283         }
 284 }
 285 EXPORT_SYMBOL(ptlrpc_abort_bulk);
 286 #endif /* HAVE_SERVER_SUPPORT */
 287
 288 /**
 289  * Register bulk at the sender for later transfer.
 290  * Returns 0 on success or error code.
 291  */
 292 int ptlrpc_register_bulk(struct ptlrpc_request *req)
 293 {
 294         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 295         lnet_process_id_t peer;
 296         int rc = 0;
 297         int rc2;
 298         int posted_md;
 299         int total_md;
 300         __u64 xid;
 301         lnet_handle_me_t  me_h;
 302         lnet_md_t         md;
 303         ENTRY;
 304
 305         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_GET_NET))
 306                 RETURN(0);
 307
 308         /* NB no locking required until desc is on the network */
 309         LASSERT(desc->bd_nob > 0);
 310         LASSERT(desc->bd_md_count == 0);
 311         LASSERT(desc->bd_md_max_brw <= PTLRPC_BULK_OPS_COUNT);
 312         LASSERT(desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
 313         LASSERT(desc->bd_req != NULL);
 314         LASSERT(desc->bd_type == BULK_PUT_SINK ||
 315                 desc->bd_type == BULK_GET_SOURCE);
 316
 317         /* cleanup the state of the bulk for it will be reused */
 318         if (req->rq_resend || req->rq_send_state == LUSTRE_IMP_REPLAY)
 319                 desc->bd_nob_transferred = 0;
 320         else
 321                 LASSERT(desc->bd_nob_transferred == 0);
 322
 323         desc->bd_failure = 0;
 324
 325         peer = desc->bd_import->imp_connection->c_peer;
 326
 327         LASSERT(desc->bd_cbid.cbid_fn == client_bulk_callback);
 328         LASSERT(desc->bd_cbid.cbid_arg == desc);
 329
 330         /* An XID is only used for a single request from the client.
 331          * For retried bulk transfers, a new XID will be allocated in
 332          * in ptlrpc_check_set() if it needs to be resent, so it is not
 333          * using the same RDMA match bits after an error.
 334          *
 335          * For multi-bulk RPCs, rq_xid is the last XID needed for bulks. The
 336          * first bulk XID is power-of-two aligned before rq_xid. LU-1431 */
 337         xid = req->rq_xid & ~((__u64)desc->bd_md_max_brw - 1);
 338         LASSERTF(!(desc->bd_registered &&
 339                    req->rq_send_state != LUSTRE_IMP_REPLAY) ||
 340                  xid != desc->bd_last_xid,
 341                  "registered: %d  rq_xid: "LPU64" bd_last_xid: "LPU64"\n",
 342                  desc->bd_registered, xid, desc->bd_last_xid);
 343
 344         total_md = (desc->bd_iov_count + LNET_MAX_IOV - 1) / LNET_MAX_IOV;
 345         desc->bd_registered = 1;
 346         desc->bd_last_xid = xid;
 347         desc->bd_md_count = total_md;
 348         md.user_ptr = &desc->bd_cbid;
 349         md.eq_handle = ptlrpc_eq_h;
 350         md.threshold = 1;                       /* PUT or GET */
 351
 352         for (posted_md = 0; posted_md < total_md; posted_md++, xid++) {
 353                 md.options = PTLRPC_MD_OPTIONS |
 354                              ((desc->bd_type == BULK_GET_SOURCE) ?
 355                               LNET_MD_OP_GET : LNET_MD_OP_PUT);
 356                 ptlrpc_fill_bulk_md(&md, desc, posted_md);
 357
 358                 rc = LNetMEAttach(desc->bd_portal, peer, xid, 0,
 359                                   LNET_UNLINK, LNET_INS_AFTER, &me_h);
 360                 if (rc != 0) {
 361                         CERROR("%s: LNetMEAttach failed x"LPU64"/%d: rc = %d\n",
 362                                desc->bd_import->imp_obd->obd_name, xid,
 363                                posted_md, rc);
 364                         break;
 365                 }
 366
 367                 /* About to let the network at it... */
 368                 rc = LNetMDAttach(me_h, md, LNET_UNLINK,
 369                                   &desc->bd_mds[posted_md]);
 370                 if (rc != 0) {
 371                         CERROR("%s: LNetMDAttach failed x"LPU64"/%d: rc = %d\n",
 372                                desc->bd_import->imp_obd->obd_name, xid,
 373                                posted_md, rc);
 374                         rc2 = LNetMEUnlink(me_h);
 375                         LASSERT(rc2 == 0);
 376                         break;
 377                 }
 378         }
 379
 380         if (rc != 0) {
 381                 LASSERT(rc == -ENOMEM);
 382                 spin_lock(&desc->bd_lock);
 383                 desc->bd_md_count -= total_md - posted_md;
 384                 spin_unlock(&desc->bd_lock);
 385                 LASSERT(desc->bd_md_count >= 0);
 386                 mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 387                 req->rq_status = -ENOMEM;
 388                 RETURN(-ENOMEM);
 389         }
 390
 391         /* Set rq_xid to matchbits of the final bulk so that server can
 392          * infer the number of bulks that were prepared */
 393         req->rq_xid = --xid;
 394         LASSERTF(desc->bd_last_xid == (req->rq_xid & PTLRPC_BULK_OPS_MASK),
 395                  "bd_last_xid = x"LPU64", rq_xid = x"LPU64"\n",
 396                  desc->bd_last_xid, req->rq_xid);
 397
 398         spin_lock(&desc->bd_lock);
 399         /* Holler if peer manages to touch buffers before he knows the xid */
 400         if (desc->bd_md_count != total_md)
 401                 CWARN("%s: Peer %s touched %d buffers while I registered\n",
 402                       desc->bd_import->imp_obd->obd_name, libcfs_id2str(peer),
 403                       total_md - desc->bd_md_count);
 404         spin_unlock(&desc->bd_lock);
 405
 406         CDEBUG(D_NET, "Setup %u bulk %s buffers: %u pages %u bytes, "
 407                "xid x"LPX64"-"LPX64", portal %u\n", desc->bd_md_count,
 408                desc->bd_type == BULK_GET_SOURCE ? "get-source" : "put-sink",
 409                desc->bd_iov_count, desc->bd_nob,
 410                desc->bd_last_xid, req->rq_xid, desc->bd_portal);
 411
 412         RETURN(0);
 413 }
 414 EXPORT_SYMBOL(ptlrpc_register_bulk);
 415
 416 /**
 417  * Disconnect a bulk desc from the network. Idempotent. Not
 418  * thread-safe (i.e. only interlocks with completion callback).
 419  * Returns 1 on success or 0 if network unregistration failed for whatever
 420  * reason.
 421  */
 422 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
 423 {
 424         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 425         struct l_wait_info       lwi;
 426         int                      rc;
 427         ENTRY;
 428
 429         LASSERT(!in_interrupt());     /* might sleep */
 430
 431         /* Let's setup deadline for reply unlink. */
 432         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
 433             async && req->rq_bulk_deadline == 0)
 434                 req->rq_bulk_deadline = cfs_time_current_sec() + LONG_UNLINK;
 435
 436         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 437                 RETURN(1);                              /* never registered */
 438
 439         LASSERT(desc->bd_req == req);  /* bd_req NULL until registered */
 440
 441         /* the unlink ensures the callback happens ASAP and is the last
 442          * one.  If it fails, it must be because completion just happened,
 443          * but we must still l_wait_event() in this case to give liblustre
 444          * a chance to run client_bulk_callback() */
 445         mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 446
 447         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 448                 RETURN(1);                              /* never registered */
 449
 450         /* Move to "Unregistering" phase as bulk was not unlinked yet. */
 451         ptlrpc_rqphase_move(req, RQ_PHASE_UNREGISTERING);
 452
 453         /* Do not wait for unlink to finish. */
 454         if (async)
 455                 RETURN(0);
 456
 457         for (;;) {
 458                 /* The wq argument is ignored by user-space wait_event macros */
 459                 wait_queue_head_t *wq = (req->rq_set != NULL) ?
 460                                         &req->rq_set->set_waitq :
 461                                         &req->rq_reply_waitq;
 462                 /* Network access will complete in finite time but the HUGE
 463                  * timeout lets us CWARN for visibility of sluggish NALs */
 464                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 465                                            cfs_time_seconds(1), NULL, NULL);
 466                 rc = l_wait_event(*wq, !ptlrpc_client_bulk_active(req), &lwi);
 467                 if (rc == 0) {
 468                         ptlrpc_rqphase_move(req, req->rq_next_phase);
 469                         RETURN(1);
 470                 }
 471
 472                 LASSERT(rc == -ETIMEDOUT);
 473                 DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p",
 474                           desc);
 475         }
 476         RETURN(0);
 477 }
 478 EXPORT_SYMBOL(ptlrpc_unregister_bulk);
 479
 480 static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags)
 481 {
 482         struct ptlrpc_service_part      *svcpt = req->rq_rqbd->rqbd_svcpt;
 483         struct ptlrpc_service           *svc = svcpt->scp_service;
 484         int service_time = max_t(int, cfs_time_current_sec() -
 485                                  req->rq_arrival_time.tv_sec, 1);
 486
 487         if (!(flags & PTLRPC_REPLY_EARLY) &&
 488             (req->rq_type != PTL_RPC_MSG_ERR) &&
 489             (req->rq_reqmsg != NULL) &&
 490             !(lustre_msg_get_flags(req->rq_reqmsg) &
 491               (MSG_RESENT | MSG_REPLAY |
 492                MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))) {
 493                 /* early replies, errors and recovery requests don't count
 494                  * toward our service time estimate */
 495                 int oldse = at_measured(&svcpt->scp_at_estimate, service_time);
 496
 497                 if (oldse != 0) {
 498                         DEBUG_REQ(D_ADAPTTO, req,
 499                                   "svc %s changed estimate from %d to %d",
 500                                   svc->srv_name, oldse,
 501                                   at_get(&svcpt->scp_at_estimate));
 502                 }
 503         }
 504         /* Report actual service time for client latency calc */
 505         lustre_msg_set_service_time(req->rq_repmsg, service_time);
 506         /* Report service time estimate for future client reqs, but report 0
 507          * (to be ignored by client) if it's a error reply during recovery.
 508          * (bz15815) */
 509         if (req->rq_type == PTL_RPC_MSG_ERR &&
 510             (req->rq_export == NULL || req->rq_export->exp_obd->obd_recovering))
 511                 lustre_msg_set_timeout(req->rq_repmsg, 0);
 512         else
 513                 lustre_msg_set_timeout(req->rq_repmsg,
 514                                        at_get(&svcpt->scp_at_estimate));
 515
 516         if (req->rq_reqmsg &&
 517             !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
 518                 CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x "
 519                        "req_flags=%#x magic=%d:%x/%x len=%d\n",
 520                        flags, lustre_msg_get_flags(req->rq_reqmsg),
 521                        lustre_msg_is_v1(req->rq_reqmsg),
 522                        lustre_msg_get_magic(req->rq_reqmsg),
 523                        lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
 524         }
 525 }
 526
 527 /**
 528  * Send request reply from request \a req reply buffer.
 529  * \a flags defines reply types
 530  * Returns 0 on sucess or error code
 531  */
 532 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
 533 {
 534         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 535         struct ptlrpc_connection  *conn;
 536         int                        rc;
 537
 538         /* We must already have a reply buffer (only ptlrpc_error() may be
 539          * called without one). The reply generated by sptlrpc layer (e.g.
 540          * error notify, etc.) might have NULL rq->reqmsg; Otherwise we must
 541          * have a request buffer which is either the actual (swabbed) incoming
 542          * request, or a saved copy if this is a req saved in
 543          * target_queue_final_reply().
 544          */
 545         LASSERT (req->rq_no_reply == 0);
 546         LASSERT (req->rq_reqbuf != NULL);
 547         LASSERT (rs != NULL);
 548         LASSERT ((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
 549         LASSERT (req->rq_repmsg != NULL);
 550         LASSERT (req->rq_repmsg == rs->rs_msg);
 551         LASSERT (rs->rs_cb_id.cbid_fn == reply_out_callback);
 552         LASSERT (rs->rs_cb_id.cbid_arg == rs);
 553
 554         /* There may be no rq_export during failover */
 555
 556         if (unlikely(req->rq_export && req->rq_export->exp_obd &&
 557                      req->rq_export->exp_obd->obd_fail)) {
 558                 /* Failed obd's only send ENODEV */
 559                 req->rq_type = PTL_RPC_MSG_ERR;
 560                 req->rq_status = -ENODEV;
 561                 CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
 562                        req->rq_export->exp_obd->obd_minor);
 563         }
 564
 565         /* In order to keep interoprability with the client (< 2.3) which
 566          * doesn't have pb_jobid in ptlrpc_body, We have to shrink the
 567          * ptlrpc_body in reply buffer to ptlrpc_body_v2, otherwise, the
 568          * reply buffer on client will be overflow.
 569          *
 570          * XXX Remove this whenver we drop the interoprability with such client.
 571          */
 572         req->rq_replen = lustre_shrink_msg(req->rq_repmsg, 0,
 573                                            sizeof(struct ptlrpc_body_v2), 1);
 574
 575         if (req->rq_type != PTL_RPC_MSG_ERR)
 576                 req->rq_type = PTL_RPC_MSG_REPLY;
 577
 578         lustre_msg_set_type(req->rq_repmsg, req->rq_type);
 579         lustre_msg_set_status(req->rq_repmsg,
 580                               ptlrpc_status_hton(req->rq_status));
 581         lustre_msg_set_opc(req->rq_repmsg,
 582                 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : 0);
 583
 584         target_pack_pool_reply(req);
 585
 586         ptlrpc_at_set_reply(req, flags);
 587
 588         if (req->rq_export == NULL || req->rq_export->exp_connection == NULL)
 589                 conn = ptlrpc_connection_get(req->rq_peer, req->rq_self, NULL);
 590         else
 591                 conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
 592
 593         if (unlikely(conn == NULL)) {
 594                 CERROR("not replying on NULL connection\n"); /* bug 9635 */
 595                 return -ENOTCONN;
 596         }
 597         ptlrpc_rs_addref(rs);                   /* +1 ref for the network */
 598
 599         rc = sptlrpc_svc_wrap_reply(req);
 600         if (unlikely(rc))
 601                 goto out;
 602
 603         req->rq_sent = cfs_time_current_sec();
 604
 605         rc = ptl_send_buf (&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
 606                            (rs->rs_difficult && !rs->rs_no_ack) ?
 607                            LNET_ACK_REQ : LNET_NOACK_REQ,
 608                            &rs->rs_cb_id, conn,
 609                            ptlrpc_req2svc(req)->srv_rep_portal,
 610                            req->rq_xid, req->rq_reply_off);
 611 out:
 612         if (unlikely(rc != 0))
 613                 ptlrpc_req_drop_rs(req);
 614         ptlrpc_connection_put(conn);
 615         return rc;
 616 }
 617 EXPORT_SYMBOL(ptlrpc_send_reply);
 618
 619 int ptlrpc_reply (struct ptlrpc_request *req)
 620 {
 621         if (req->rq_no_reply)
 622                 return 0;
 623         else
 624                 return (ptlrpc_send_reply(req, 0));
 625 }
 626 EXPORT_SYMBOL(ptlrpc_reply);
 627
 628 /**
 629  * For request \a req send an error reply back. Create empty
 630  * reply buffers if necessary.
 631  */
 632 int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult)
 633 {
 634         int rc;
 635         ENTRY;
 636
 637         if (req->rq_no_reply)
 638                 RETURN(0);
 639
 640         if (!req->rq_repmsg) {
 641                 rc = lustre_pack_reply(req, 1, NULL, NULL);
 642                 if (rc)
 643                         RETURN(rc);
 644         }
 645
 646         if (req->rq_status != -ENOSPC && req->rq_status != -EACCES &&
 647             req->rq_status != -EPERM && req->rq_status != -ENOENT &&
 648             req->rq_status != -EINPROGRESS && req->rq_status != -EDQUOT)
 649                 req->rq_type = PTL_RPC_MSG_ERR;
 650
 651         rc = ptlrpc_send_reply(req, may_be_difficult);
 652         RETURN(rc);
 653 }
 654 EXPORT_SYMBOL(ptlrpc_send_error);
 655
 656 int ptlrpc_error(struct ptlrpc_request *req)
 657 {
 658         return ptlrpc_send_error(req, 0);
 659 }
 660 EXPORT_SYMBOL(ptlrpc_error);
 661
 662 /**
 663  * Send request \a request.
 664  * if \a noreply is set, don't expect any reply back and don't set up
 665  * reply buffers.
 666  * Returns 0 on success or error code.
 667  */
 668 int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
 669 {
 670         int rc;
 671         int rc2;
 672         int mpflag = 0;
 673         struct ptlrpc_connection *connection;
 674         lnet_handle_me_t  reply_me_h;
 675         lnet_md_t         reply_md;
 676         struct obd_device *obd = request->rq_import->imp_obd;
 677         ENTRY;
 678
 679         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
 680                 RETURN(0);
 681
 682         LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
 683         LASSERT(request->rq_wait_ctx == 0);
 684
 685         /* If this is a re-transmit, we're required to have disengaged
 686          * cleanly from the previous attempt */
 687         LASSERT(!request->rq_receiving_reply);
 688         LASSERT(!((lustre_msg_get_flags(request->rq_reqmsg) & MSG_REPLAY) &&
 689                 (request->rq_import->imp_state == LUSTRE_IMP_FULL)));
 690
 691         if (unlikely(obd != NULL && obd->obd_fail)) {
 692                 CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
 693                         obd->obd_name);
 694                 /* this prevents us from waiting in ptlrpc_queue_wait */
 695                 spin_lock(&request->rq_lock);
 696                 request->rq_err = 1;
 697                 spin_unlock(&request->rq_lock);
 698                 request->rq_status = -ENODEV;
 699                 RETURN(-ENODEV);
 700         }
 701
 702         connection = request->rq_import->imp_connection;
 703
 704         lustre_msg_set_handle(request->rq_reqmsg,
 705                               &request->rq_import->imp_remote_handle);
 706         lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
 707         lustre_msg_set_conn_cnt(request->rq_reqmsg,
 708                                 request->rq_import->imp_conn_cnt);
 709         lustre_msghdr_set_flags(request->rq_reqmsg,
 710                                 request->rq_import->imp_msghdr_flags);
 711
 712         if (request->rq_resend) {
 713                 lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
 714
 715                 if (request->rq_resend_cb != NULL)
 716                         request->rq_resend_cb(request, &request->rq_async_args);
 717         }
 718         if (request->rq_memalloc)
 719                 mpflag = cfs_memory_pressure_get_and_set();
 720
 721         rc = sptlrpc_cli_wrap_request(request);
 722         if (rc)
 723                 GOTO(out, rc);
 724
 725         /* bulk register should be done after wrap_request() */
 726         if (request->rq_bulk != NULL) {
 727                 rc = ptlrpc_register_bulk (request);
 728                 if (rc != 0)
 729                         GOTO(out, rc);
 730         }
 731
 732         if (!noreply) {
 733                 LASSERT (request->rq_replen != 0);
 734                 if (request->rq_repbuf == NULL) {
 735                         LASSERT(request->rq_repdata == NULL);
 736                         LASSERT(request->rq_repmsg == NULL);
 737                         rc = sptlrpc_cli_alloc_repbuf(request,
 738                                                       request->rq_replen);
 739                         if (rc) {
 740                                 /* this prevents us from looping in
 741                                  * ptlrpc_queue_wait */
 742                                 spin_lock(&request->rq_lock);
 743                                 request->rq_err = 1;
 744                                 spin_unlock(&request->rq_lock);
 745                                 request->rq_status = rc;
 746                                 GOTO(cleanup_bulk, rc);
 747                         }
 748                 } else {
 749                         request->rq_repdata = NULL;
 750                         request->rq_repmsg = NULL;
 751                 }
 752
 753                 rc = LNetMEAttach(request->rq_reply_portal,/*XXX FIXME bug 249*/
 754                                   connection->c_peer, request->rq_xid, 0,
 755                                   LNET_UNLINK, LNET_INS_AFTER, &reply_me_h);
 756                 if (rc != 0) {
 757                         CERROR("LNetMEAttach failed: %d\n", rc);
 758                         LASSERT (rc == -ENOMEM);
 759                         GOTO(cleanup_bulk, rc = -ENOMEM);
 760                 }
 761         }
 762
 763         spin_lock(&request->rq_lock);
 764         /* If the MD attach succeeds, there _will_ be a reply_in callback */
 765         request->rq_receiving_reply = !noreply;
 766         request->rq_req_unlink = 1;
 767         /* We are responsible for unlinking the reply buffer */
 768         request->rq_reply_unlink = !noreply;
 769         /* Clear any flags that may be present from previous sends. */
 770         request->rq_replied = 0;
 771         request->rq_err = 0;
 772         request->rq_timedout = 0;
 773         request->rq_net_err = 0;
 774         request->rq_resend = 0;
 775         request->rq_restart = 0;
 776         request->rq_reply_truncate = 0;
 777         spin_unlock(&request->rq_lock);
 778
 779         if (!noreply) {
 780                 reply_md.start     = request->rq_repbuf;
 781                 reply_md.length    = request->rq_repbuf_len;
 782                 /* Allow multiple early replies */
 783                 reply_md.threshold = LNET_MD_THRESH_INF;
 784                 /* Manage remote for early replies */
 785                 reply_md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
 786                         LNET_MD_MANAGE_REMOTE |
 787                         LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */;
 788                 reply_md.user_ptr  = &request->rq_reply_cbid;
 789                 reply_md.eq_handle = ptlrpc_eq_h;
 790
 791                 /* We must see the unlink callback to unset rq_reply_unlink,
 792                    so we can't auto-unlink */
 793                 rc = LNetMDAttach(reply_me_h, reply_md, LNET_RETAIN,
 794                                   &request->rq_reply_md_h);
 795                 if (rc != 0) {
 796                         CERROR("LNetMDAttach failed: %d\n", rc);
 797                         LASSERT (rc == -ENOMEM);
 798                         spin_lock(&request->rq_lock);
 799                         /* ...but the MD attach didn't succeed... */
 800                         request->rq_receiving_reply = 0;
 801                         spin_unlock(&request->rq_lock);
 802                         GOTO(cleanup_me, rc = -ENOMEM);
 803                 }
 804
 805                 CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid "LPU64
 806                        ", portal %u\n",
 807                        request->rq_repbuf_len, request->rq_xid,
 808                        request->rq_reply_portal);
 809         }
 810
 811         /* add references on request for request_out_callback */
 812         ptlrpc_request_addref(request);
 813         if (obd != NULL && obd->obd_svc_stats != NULL)
 814                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
 815                         atomic_read(&request->rq_import->imp_inflight));
 816
 817         OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
 818
 819         do_gettimeofday(&request->rq_sent_tv);
 820         request->rq_sent = cfs_time_current_sec();
 821         /* We give the server rq_timeout secs to process the req, and
 822            add the network latency for our local timeout. */
 823         request->rq_deadline = request->rq_sent + request->rq_timeout +
 824                 ptlrpc_at_get_net_latency(request);
 825
 826         ptlrpc_pinger_sending_on_import(request->rq_import);
 827
 828         DEBUG_REQ(D_INFO, request, "send flg=%x",
 829                   lustre_msg_get_flags(request->rq_reqmsg));
 830         rc = ptl_send_buf(&request->rq_req_md_h,
 831                           request->rq_reqbuf, request->rq_reqdata_len,
 832                           LNET_NOACK_REQ, &request->rq_req_cbid,
 833                           connection,
 834                           request->rq_request_portal,
 835                           request->rq_xid, 0);
 836         if (rc == 0)
 837                 GOTO(out, rc);
 838
 839         ptlrpc_req_finished(request);
 840         if (noreply)
 841                 GOTO(out, rc);
 842
 843  cleanup_me:
 844         /* MEUnlink is safe; the PUT didn't even get off the ground, and
 845          * nobody apart from the PUT's target has the right nid+XID to
 846          * access the reply buffer. */
 847         rc2 = LNetMEUnlink(reply_me_h);
 848         LASSERT (rc2 == 0);
 849         /* UNLINKED callback called synchronously */
 850         LASSERT(!request->rq_receiving_reply);
 851
 852  cleanup_bulk:
 853         /* We do sync unlink here as there was no real transfer here so
 854          * the chance to have long unlink to sluggish net is smaller here. */
 855         ptlrpc_unregister_bulk(request, 0);
 856  out:
 857         if (request->rq_memalloc)
 858                 cfs_memory_pressure_restore(mpflag);
 859         return rc;
 860 }
 861 EXPORT_SYMBOL(ptl_send_rpc);
 862
 863 /**
 864  * Register request buffer descriptor for request receiving.
 865  */
 866 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
 867 {
 868         struct ptlrpc_service     *service = rqbd->rqbd_svcpt->scp_service;
 869         static lnet_process_id_t  match_id = {LNET_NID_ANY, LNET_PID_ANY};
 870         int                       rc;
 871         lnet_md_t                 md;
 872         lnet_handle_me_t          me_h;
 873
 874         CDEBUG(D_NET, "LNetMEAttach: portal %d\n",
 875                service->srv_req_portal);
 876
 877         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD))
 878                 return (-ENOMEM);
 879
 880         /* NB: CPT affinity service should use new LNet flag LNET_INS_LOCAL,
 881          * which means buffer can only be attached on local CPT, and LND
 882          * threads can find it by grabbing a local lock */
 883         rc = LNetMEAttach(service->srv_req_portal,
 884                           match_id, 0, ~0, LNET_UNLINK,
 885                           rqbd->rqbd_svcpt->scp_cpt >= 0 ?
 886                           LNET_INS_LOCAL : LNET_INS_AFTER, &me_h);
 887         if (rc != 0) {
 888                 CERROR("LNetMEAttach failed: %d\n", rc);
 889                 return (-ENOMEM);
 890         }
 891
 892         LASSERT(rqbd->rqbd_refcount == 0);
 893         rqbd->rqbd_refcount = 1;
 894
 895         md.start     = rqbd->rqbd_buffer;
 896         md.length    = service->srv_buf_size;
 897         md.max_size  = service->srv_max_req_size;
 898         md.threshold = LNET_MD_THRESH_INF;
 899         md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
 900         md.user_ptr  = &rqbd->rqbd_cbid;
 901         md.eq_handle = ptlrpc_eq_h;
 902
 903         rc = LNetMDAttach(me_h, md, LNET_UNLINK, &rqbd->rqbd_md_h);
 904         if (rc == 0)
 905                 return (0);
 906
 907         CERROR("LNetMDAttach failed: %d; \n", rc);
 908         LASSERT (rc == -ENOMEM);
 909         rc = LNetMEUnlink (me_h);
 910         LASSERT (rc == 0);
 911         rqbd->rqbd_refcount = 0;
 912
 913         return (-ENOMEM);
 914 }